摘要 |
PROBLEM TO BE SOLVED: To properly update a weighting matrix even if a record number which is a training example for updating the weighting matrix constituting an estimation model is biased depending on classes.SOLUTION: An estimation model creation unit 4 sets one or more classes as objects, selects, from a training data DB 3, a weight vector (w) composed of a weight parameter related to a non-objective class (i) about each objective class (c), a weight vector (w) composed of a weight parameter value related to the objective class (c) from an internal product of a feature vector (x) composed of each feature in one record of the training data DB, and a record which maximizes a maximum loss obtained by subtracting the internal product of the feature vector x, and updates the weight matrix by using a training example aggregate including the selected record. |